app/test-compress-perf/main.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2018 Intel Corporation
   3  */
   4
   5 #include <signal.h>
   6 #include <sys/types.h>
   7 #include <unistd.h>
   8
   9 #include <rte_malloc.h>
  10 #include <rte_eal.h>
  11 #include <rte_log.h>
  12 #include <rte_compressdev.h>
  13
  14 #include "comp_perf_options.h"
  15 #include "comp_perf_test_verify.h"
  16 #include "comp_perf_test_benchmark.h"
  17 #include "comp_perf.h"
  18 #include "comp_perf_test_common.h"
  19
  20 #define NUM_MAX_XFORMS 16
  21 #define NUM_MAX_INFLIGHT_OPS 512
  22
  23 __extension__
  24 const char *cperf_test_type_strs[] = {
  25         [CPERF_TEST_TYPE_BENCHMARK] = "benchmark",
  26         [CPERF_TEST_TYPE_VERIFY] = "verify"
  27 };
  28
  29 __extension__
  30 static const struct cperf_test cperf_testmap[] = {
  31         [CPERF_TEST_TYPE_BENCHMARK] = {
  32                         cperf_benchmark_test_constructor,
  33                         cperf_benchmark_test_runner,
  34                         cperf_benchmark_test_destructor
  35         },
  36         [CPERF_TEST_TYPE_VERIFY] = {
  37                         cperf_verify_test_constructor,
  38                         cperf_verify_test_runner,
  39                         cperf_verify_test_destructor
  40         }
  41 };
  42
  43 static struct comp_test_data *test_data;
  44
  45 static int
  46 comp_perf_check_capabilities(struct comp_test_data *test_data, uint8_t cdev_id)
  47 {
  48         const struct rte_compressdev_capabilities *cap;
  49
  50         cap = rte_compressdev_capability_get(cdev_id,
  51                                              RTE_COMP_ALGO_DEFLATE);
  52
  53         if (cap == NULL) {
  54                 RTE_LOG(ERR, USER1,
  55                         "Compress device does not support DEFLATE\n");
  56                 return -1;
  57         }
  58
  59         uint64_t comp_flags = cap->comp_feature_flags;
  60
  61         /* Huffman enconding */
  62         if (test_data->huffman_enc == RTE_COMP_HUFFMAN_FIXED &&
  63                         (comp_flags & RTE_COMP_FF_HUFFMAN_FIXED) == 0) {
  64                 RTE_LOG(ERR, USER1,
  65                         "Compress device does not supported Fixed Huffman\n");
  66                 return -1;
  67         }
  68
  69         if (test_data->huffman_enc == RTE_COMP_HUFFMAN_DYNAMIC &&
  70                         (comp_flags & RTE_COMP_FF_HUFFMAN_DYNAMIC) == 0) {
  71                 RTE_LOG(ERR, USER1,
  72                         "Compress device does not supported Dynamic Huffman\n");
  73                 return -1;
  74         }
  75
  76         /* Window size */
  77         if (test_data->window_sz != -1) {
  78                 if (param_range_check(test_data->window_sz, &cap->window_size)
  79                                 < 0) {
  80                         RTE_LOG(ERR, USER1,
  81                                 "Compress device does not support "
  82                                 "this window size\n");
  83                         return -1;
  84                 }
  85         } else
  86                 /* Set window size to PMD maximum if none was specified */
  87                 test_data->window_sz = cap->window_size.max;
  88
  89         /* Check if chained mbufs is supported */
  90         if (test_data->max_sgl_segs > 1  &&
  91                         (comp_flags & RTE_COMP_FF_OOP_SGL_IN_SGL_OUT) == 0) {
  92                 RTE_LOG(INFO, USER1, "Compress device does not support "
  93                                 "chained mbufs. Max SGL segments set to 1\n");
  94                 test_data->max_sgl_segs = 1;
  95         }
  96
  97         /* Level 0 support */
  98         if (test_data->level_lst.min == 0 &&
  99                         (comp_flags & RTE_COMP_FF_NONCOMPRESSED_BLOCKS) == 0) {
 100                 RTE_LOG(ERR, USER1, "Compress device does not support "
 101                                 "level 0 (no compression)\n");
 102                 return -1;
 103         }
 104
 105         return 0;
 106 }
 107
 108 static int
 109 comp_perf_initialize_compressdev(struct comp_test_data *test_data,
 110                                  uint8_t *enabled_cdevs)
 111 {
 112         uint8_t enabled_cdev_count, nb_lcores, cdev_id;
 113         unsigned int i, j;
 114         int ret;
 115
 116         enabled_cdev_count = rte_compressdev_devices_get(test_data->driver_name,
 117                         enabled_cdevs, RTE_COMPRESS_MAX_DEVS);
 118         if (enabled_cdev_count == 0) {
 119                 RTE_LOG(ERR, USER1, "No compress devices type %s available\n",
 120                                 test_data->driver_name);
 121                 return -EINVAL;
 122         }
 123
 124         nb_lcores = rte_lcore_count() - 1;
 125         /*
 126          * Use fewer devices,
 127          * if there are more available than cores.
 128          */
 129         if (enabled_cdev_count > nb_lcores) {
 130                 enabled_cdev_count = nb_lcores;
 131                 RTE_LOG(INFO, USER1,
 132                         " There's more available devices than cores!"
 133                         " The number of devices has been aligned to %d cores\n",
 134                         nb_lcores);
 135         }
 136
 137         /*
 138          * Calculate number of needed queue pairs, based on the amount
 139          * of available number of logical cores and compression devices.
 140          * For instance, if there are 4 cores and 2 compression devices,
 141          * 2 queue pairs will be set up per device.
 142          * One queue pair per one core.
 143          * if e.g.: there're 3 cores and 2 compression devices,
 144          * 2 queue pairs will be set up per device but one queue pair
 145          * will left unused in the last one device
 146          */
 147         test_data->nb_qps = (nb_lcores % enabled_cdev_count) ?
 148                                 (nb_lcores / enabled_cdev_count) + 1 :
 149                                 nb_lcores / enabled_cdev_count;
 150
 151         for (i = 0; i < enabled_cdev_count &&
 152                         i < RTE_COMPRESS_MAX_DEVS; i++,
 153                                         nb_lcores -= test_data->nb_qps) {
 154                 cdev_id = enabled_cdevs[i];
 155
 156                 struct rte_compressdev_info cdev_info;
 157                 uint8_t socket_id = rte_compressdev_socket_id(cdev_id);
 158
 159                 rte_compressdev_info_get(cdev_id, &cdev_info);
 160                 if (cdev_info.max_nb_queue_pairs &&
 161                         test_data->nb_qps > cdev_info.max_nb_queue_pairs) {
 162                         RTE_LOG(ERR, USER1,
 163                                 "Number of needed queue pairs is higher "
 164                                 "than the maximum number of queue pairs "
 165                                 "per device.\n");
 166                         RTE_LOG(ERR, USER1,
 167                                 "Lower the number of cores or increase "
 168                                 "the number of crypto devices\n");
 169                         return -EINVAL;
 170                 }
 171
 172                 if (comp_perf_check_capabilities(test_data, cdev_id) < 0)
 173                         return -EINVAL;
 174
 175                 /* Configure compressdev */
 176                 struct rte_compressdev_config config = {
 177                         .socket_id = socket_id,
 178                         .nb_queue_pairs = nb_lcores > test_data->nb_qps
 179                                         ? test_data->nb_qps : nb_lcores,
 180                         .max_nb_priv_xforms = NUM_MAX_XFORMS,
 181                         .max_nb_streams = 0
 182                 };
 183
 184                 if (rte_compressdev_configure(cdev_id, &config) < 0) {
 185                         RTE_LOG(ERR, USER1, "Device configuration failed\n");
 186                         return -EINVAL;
 187                 }
 188
 189                 for (j = 0; j < test_data->nb_qps; j++) {
 190                         ret = rte_compressdev_queue_pair_setup(cdev_id, j,
 191                                         NUM_MAX_INFLIGHT_OPS, socket_id);
 192                         if (ret < 0) {
 193                                 RTE_LOG(ERR, USER1,
 194                               "Failed to setup queue pair %u on compressdev %u",
 195                                         j, cdev_id);
 196                                 return -EINVAL;
 197                         }
 198                 }
 199
 200                 ret = rte_compressdev_start(cdev_id);
 201                 if (ret < 0) {
 202                         RTE_LOG(ERR, USER1,
 203                                 "Failed to start device %u: error %d\n",
 204                                 cdev_id, ret);
 205                         return -EPERM;
 206                 }
 207         }
 208
 209         return enabled_cdev_count;
 210 }
 211
 212 static int
 213 comp_perf_dump_input_data(struct comp_test_data *test_data)
 214 {
 215         FILE *f = fopen(test_data->input_file, "r");
 216         int ret = -1;
 217
 218         if (f == NULL) {
 219                 RTE_LOG(ERR, USER1, "Input file could not be opened\n");
 220                 return -1;
 221         }
 222
 223         if (fseek(f, 0, SEEK_END) != 0) {
 224                 RTE_LOG(ERR, USER1, "Size of input could not be calculated\n");
 225                 goto end;
 226         }
 227         size_t actual_file_sz = ftell(f);
 228         /* If extended input data size has not been set,
 229          * input data size = file size
 230          */
 231
 232         if (test_data->input_data_sz == 0)
 233                 test_data->input_data_sz = actual_file_sz;
 234
 235         if (test_data->input_data_sz <= 0 || actual_file_sz <= 0 ||
 236                         fseek(f, 0, SEEK_SET) != 0) {
 237                 RTE_LOG(ERR, USER1, "Size of input could not be calculated\n");
 238                 goto end;
 239         }
 240
 241         test_data->input_data = rte_zmalloc_socket(NULL,
 242                                 test_data->input_data_sz, 0, rte_socket_id());
 243
 244         if (test_data->input_data == NULL) {
 245                 RTE_LOG(ERR, USER1, "Memory to hold the data from the input "
 246                                 "file could not be allocated\n");
 247                 goto end;
 248         }
 249
 250         size_t remaining_data = test_data->input_data_sz;
 251         uint8_t *data = test_data->input_data;
 252
 253         while (remaining_data > 0) {
 254                 size_t data_to_read = RTE_MIN(remaining_data, actual_file_sz);
 255
 256                 if (fread(data, data_to_read, 1, f) != 1) {
 257                         RTE_LOG(ERR, USER1, "Input file could not be read\n");
 258                         goto end;
 259                 }
 260                 if (fseek(f, 0, SEEK_SET) != 0) {
 261                         RTE_LOG(ERR, USER1,
 262                                 "Size of input could not be calculated\n");
 263                         goto end;
 264                 }
 265                 remaining_data -= data_to_read;
 266                 data += data_to_read;
 267         }
 268
 269         if (test_data->input_data_sz > actual_file_sz)
 270                 RTE_LOG(INFO, USER1,
 271                   "%zu bytes read from file %s, extending the file %.2f times\n",
 272                         test_data->input_data_sz, test_data->input_file,
 273                         (double)test_data->input_data_sz/actual_file_sz);
 274         else
 275                 RTE_LOG(INFO, USER1,
 276                         "%zu bytes read from file %s\n",
 277                         test_data->input_data_sz, test_data->input_file);
 278
 279         ret = 0;
 280
 281 end:
 282         fclose(f);
 283         return ret;
 284 }
 285
 286 static void
 287 comp_perf_cleanup_on_signal(int signalNumber __rte_unused)
 288 {
 289         test_data->perf_comp_force_stop = 1;
 290 }
 291
 292 static void
 293 comp_perf_register_cleanup_on_signal(void)
 294 {
 295         signal(SIGTERM, comp_perf_cleanup_on_signal);
 296         signal(SIGINT, comp_perf_cleanup_on_signal);
 297 }
 298
 299 int
 300 main(int argc, char **argv)
 301 {
 302         uint8_t level_idx = 0;
 303         int ret, i;
 304         void *ctx[RTE_MAX_LCORE] = {};
 305         uint8_t enabled_cdevs[RTE_COMPRESS_MAX_DEVS];
 306         int nb_compressdevs = 0;
 307         uint16_t total_nb_qps = 0;
 308         uint8_t cdev_id;
 309         uint32_t lcore_id;
 310
 311         /* Initialise DPDK EAL */
 312         ret = rte_eal_init(argc, argv);
 313         if (ret < 0)
 314                 rte_exit(EXIT_FAILURE, "Invalid EAL arguments!\n");
 315         argc -= ret;
 316         argv += ret;
 317
 318         test_data = rte_zmalloc_socket(NULL, sizeof(struct comp_test_data),
 319                                         0, rte_socket_id());
 320
 321         if (test_data == NULL)
 322                 rte_exit(EXIT_FAILURE, "Cannot reserve memory in socket %d\n",
 323                                 rte_socket_id());
 324
 325         comp_perf_register_cleanup_on_signal();
 326
 327         ret = EXIT_SUCCESS;
 328         test_data->cleanup = ST_TEST_DATA;
 329         comp_perf_options_default(test_data);
 330
 331         if (comp_perf_options_parse(test_data, argc, argv) < 0) {
 332                 RTE_LOG(ERR, USER1,
 333                         "Parsing one or more user options failed\n");
 334                 ret = EXIT_FAILURE;
 335                 goto end;
 336         }
 337
 338         if (comp_perf_options_check(test_data) < 0) {
 339                 ret = EXIT_FAILURE;
 340                 goto end;
 341         }
 342
 343         nb_compressdevs =
 344                 comp_perf_initialize_compressdev(test_data, enabled_cdevs);
 345
 346         if (nb_compressdevs < 1) {
 347                 ret = EXIT_FAILURE;
 348                 goto end;
 349         }
 350
 351         test_data->cleanup = ST_COMPDEV;
 352         if (comp_perf_dump_input_data(test_data) < 0) {
 353                 ret = EXIT_FAILURE;
 354                 goto end;
 355         }
 356
 357         test_data->cleanup = ST_INPUT_DATA;
 358
 359         if (test_data->level_lst.inc != 0)
 360                 test_data->level = test_data->level_lst.min;
 361         else
 362                 test_data->level = test_data->level_lst.list[0];
 363
 364         printf("App uses socket: %u\n", rte_socket_id());
 365         printf("Burst size = %u\n", test_data->burst_sz);
 366         printf("File size = %zu\n", test_data->input_data_sz);
 367
 368         test_data->cleanup = ST_DURING_TEST;
 369         total_nb_qps = nb_compressdevs * test_data->nb_qps;
 370
 371         i = 0;
 372         uint8_t qp_id = 0, cdev_index = 0;
 373
 374         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 375
 376                 if (i == total_nb_qps)
 377                         break;
 378
 379                 cdev_id = enabled_cdevs[cdev_index];
 380                 ctx[i] = cperf_testmap[test_data->test].constructor(
 381                                                         cdev_id, qp_id,
 382                                                         test_data);
 383                 if (ctx[i] == NULL) {
 384                         RTE_LOG(ERR, USER1, "Test run constructor failed\n");
 385                         goto end;
 386                 }
 387                 qp_id = (qp_id + 1) % test_data->nb_qps;
 388                 if (qp_id == 0)
 389                         cdev_index++;
 390                 i++;
 391         }
 392
 393         while (test_data->level <= test_data->level_lst.max) {
 394
 395                 i = 0;
 396                 RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 397
 398                         if (i == total_nb_qps)
 399                                 break;
 400
 401                         rte_eal_remote_launch(
 402                                         cperf_testmap[test_data->test].runner,
 403                                         ctx[i], lcore_id);
 404                         i++;
 405                 }
 406                 i = 0;
 407                 RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 408
 409                         if (i == total_nb_qps)
 410                                 break;
 411                         ret |= rte_eal_wait_lcore(lcore_id);
 412                         i++;
 413                 }
 414
 415                 if (ret != EXIT_SUCCESS)
 416                         break;
 417
 418                 if (test_data->level_lst.inc != 0)
 419                         test_data->level += test_data->level_lst.inc;
 420                 else {
 421                         if (++level_idx == test_data->level_lst.count)
 422                                 break;
 423                         test_data->level = test_data->level_lst.list[level_idx];
 424                 }
 425         }
 426
 427 end:
 428         switch (test_data->cleanup) {
 429
 430         case ST_DURING_TEST:
 431                 i = 0;
 432                 RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 433                         if (i == total_nb_qps)
 434                                 break;
 435
 436                         if (ctx[i] && cperf_testmap[test_data->test].destructor)
 437                                 cperf_testmap[test_data->test].destructor(
 438                                                                         ctx[i]);
 439                         i++;
 440                 }
 441                 /* fallthrough */
 442         case ST_INPUT_DATA:
 443                 rte_free(test_data->input_data);
 444                 /* fallthrough */
 445         case ST_COMPDEV:
 446                 for (i = 0; i < nb_compressdevs &&
 447                      i < RTE_COMPRESS_MAX_DEVS; i++) {
 448                         rte_compressdev_stop(enabled_cdevs[i]);
 449                         rte_compressdev_close(enabled_cdevs[i]);
 450                 }
 451                 /* fallthrough */
 452         case ST_TEST_DATA:
 453                 rte_free(test_data);
 454                 /* fallthrough */
 455         case ST_CLEAR:
 456         default:
 457                 i = rte_eal_cleanup();
 458                 if (i) {
 459                         RTE_LOG(ERR, USER1,
 460                                 "Error from rte_eal_cleanup(), %d\n", i);
 461                         ret = i;
 462                 }
 463                 break;
 464         }
 465         return ret;
 466 }
 467
 468 __rte_weak void *
 469 cperf_benchmark_test_constructor(uint8_t dev_id __rte_unused,
 470                                  uint16_t qp_id __rte_unused,
 471                                  struct comp_test_data *options __rte_unused)
 472 {
 473         RTE_LOG(INFO, USER1, "Benchmark test is not supported yet\n");
 474         return NULL;
 475 }
 476
 477 __rte_weak void
 478 cperf_benchmark_test_destructor(void *arg __rte_unused)
 479 {
 480
 481 }
 482
 483 __rte_weak int
 484 cperf_benchmark_test_runner(void *test_ctx __rte_unused)
 485 {
 486         return 0;
 487 }
 488 __rte_weak void *
 489 cperf_verify_test_constructor(uint8_t dev_id __rte_unused,
 490                                  uint16_t qp_id __rte_unused,
 491                                  struct comp_test_data *options __rte_unused)
 492 {
 493         RTE_LOG(INFO, USER1, "Verify test is not supported yet\n");
 494         return NULL;
 495 }
 496
 497 __rte_weak void
 498 cperf_verify_test_destructor(void *arg __rte_unused)
 499 {
 500
 501 }
 502
 503 __rte_weak int
 504 cperf_verify_test_runner(void *test_ctx __rte_unused)
 505 {
 506         return 0;
 507 }